I feel that they won't be trying to generate novel responses from the model, but rather take knowledge graph + relevant data from the first few responses and ask the model to summarise that/change into an answer which humans find appealing.

That way you don't have to rely on the model to remember stuff, it can access all required information through attention.